DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claims 1-14 are presented for examination.
Response to Amendment
Most, but not all, of the prior objections to the specification, drawings, and claims have been obviated by the amendment. To the extent that an objection appears in both this Office Action and the previous Office Action it is maintained, otherwise, it is withdrawn. The prior rejections under 35 U.S.C. 112(b) have been obviated by the amendment. Accordingly, these rejections are withdrawn.
Specification
The disclosure is objected to because of the following grammatical/typographical informalities:
Page 1, line 7: “using semi-supervised retriever” should read “using a semi-supervised retriever”
Page 7, line 18: “using self-supervised retriever” should read “using a self-supervised retriever”
Page 8, line 30: "the .computer" should read " the computer"
Page 15, line 14: "referred as context a embedding" should read "referred to as a context embedding"
Page 15, line 26: "positive query embeddings" should read "positive query embedding"
Page 16, line 4: "negative context embedding" should read "negative context embeddings"
Page 16, line 12: “positive context embedding” should read “positive context embeddings”
Page 20, line 16: “user selected shows” should read “user selected shoes”
Page 20, lines 28-29: “medial record” should read “medical record” and “input query” should read “inputs a query”
Page 21, line 1: “looking for managing” should read “looking to manage”
Page 22, line 10: “for large language model” should read “for large language models”
Claim Objections
Claims 4 and 13 are objected to because of the following informalities:
“a plurality of vector embeddings that includes multimodal vector” is grammatically incorrect. It appears that this should read “a plurality of vector embeddings that include a multimodal vector”
Appropriate correction is required.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-14 are rejected under 35 U.S.C. 101 because they are directed to an abstract idea without significantly more. The analysis of the claims will follow the 2019 Revised Patent Subject Matter Eligibility Guidance (“2019 PEG”).
Claim 1
Step 1: The claim recites a method and therefore is directed to the statutory category of processes.
Step 2A Prong 1: The claim recites, inter alia:
“retrieving a response from one or more data sources based on the input text…”; This limitation encompasses mentally retrieving a response from one or more data sources based on the input text, such as by reviewing one or more data sources and mentally choosing a response from a data source based on the input text.
“determining… a context for each unlabelled text in the plurality of unlabelled texts”; This limitation encompasses mentally determining a context for each unlabelled text, such as by mentally determining the subject matter of the unlabelled text.
“performing an augmentation operation on each unlabelled text and the context corresponding to each unlabelled text”; This limitation encompasses mentally performing an augmentation operation on each unlabelled text and the context corresponding to each unlabelled text, such as by mentally generating additional texts and contexts to augment each unlabelled text and corresponding context.
“generating a plurality of positive and negative query-context pairs for each unlabelled text and the corresponding context based on the augmentation operation…”; This limitation encompasses mentally generating a plurality of positive and negative query-context pairs for the unlabelled text and the corresponding context based on the augmentation operation, such as by mentally pairing positive queries with positive contexts and negative queries with negative contexts.
Step 2A Prong 2: This judicial exception is not integrated into a practical application. The claim further recites “receiving an input text indicating a query made by a user”, however this limitation amounts to the insignificant extra-solution activity of mere data gathering (MPEP 2106.05(g)). The claim additionally recites “providing a plurality of unlabelled texts as training input data to the self-supervised Large Language retrieval model”, that the response is retrieved “using a self-supervised Large Language retrieval model,” that the context is determined “using one or more Artificial Intelligence (AI) techniques,” and that the plurality of positive and negative query-context pairs are generated “using the self-supervised Large Language retrieval model,” however, these limitations amount to mere instructions to apply an exception on a generic computer programmed with a generic class of computer algorithms (MPEP 2106.05(f)). The claim further recites “wherein the self-supervised Large Language retrieval model is pretrained” and “configuring the self-supervised Large Language retrieval model to retrieve the response from the one or more data sources in response to each unlabelled text,” however, these limitations merely generally link the use of the judicial exception to the technological environment of model pre-training (MPEP 2106.05(h)).
Step 2B: The claim does not contain significantly more than the judicial exception. The “receiving an input text” limitation, in addition to being insignificant extra-solution activity, is also directed to the well-understood, routine, and conventional activity of receiving or transmitting data over a network (MPEP 2106.05(d)(II)(i) buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355, 112 USPQ2d 1093, 1096 (Fed. Cir. 2014) (computer receives and sends information over a network)). Otherwise, the analysis at this step mirrors that of step 2A prong 2. As an ordered whole, the claim is directed to a mentally performable process of determining a context for unlabelled texts, performing an augmentation operation, and generating positive and negative query-context pairs for retrieving a response from one or more data sources. Nothing in the claim provides significantly more than this. As such, the claim is not patent eligible.
Claim 2
Step 1: A process, as above.
Step 2A Prong 1: The claim recites, inter alia:
“…retrieve the response from the one or more data sources based on the labelled query-context pairs”; This limitation encompasses mentally retrieving a response from one or more data sources based on the labelled query-context pairs, such as by reviewing one or more data sources and mentally choosing a response from a data source based on the labelled query-context pairs.
Step 2A Prong 2: This judicial exception is not integrated into a practical application. The claim further recites “receiving a plurality of test query-context representations such that each test query-context representation indicates a test query and a corresponding context of the test query,” however this limitation amounts to the insignificant extra-solution activity of mere data gathering (MPEP 2106.05(g)). The claim additionally recites “training the self-supervised Large Language retrieval model,” “providing labelled query-context pairs to the self-supervised Large Language retrieval model,” and “optimizing the self-supervised Large Language retrieval model to… [perform the judicial exception],” however, these limitations merely generally link the use of the judicial exception to the technological environment of model training (MPEP 2106.05(h)).
Step 2B: The claim does not contain significantly more than the judicial exception. The “receive a plurality of test query-context representations” limitation, in addition to being insignificant extra-solution activity, are also directed to the well-understood, routine, and conventional activity of receiving or transmitting data over a network (MPEP 2106.05(d)(II)(i) buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355, 112 USPQ2d 1093, 1096 (Fed. Cir. 2014) (computer receives and sends information over a network)). Otherwise, the analysis at this step mirrors that of step 2A prong 2.
Claim 3
Step 1: A process, as above.
Step 2A Prong 1: The claim recites, inter alia:
“evaluating a quadruplet loss function by evaluating a loss between the response retrieved by the self-supervised Large Language retrieval model and an expected response, wherein the quadruplet loss function is based on the positive and negative query-context pairs for each of the unlabelled text and the corresponding context”; This limitation encompasses mentally evaluating a quadruplet loss function by evaluating a loss between the response retrieved by the self-supervised Large Language retrieval model and an expected response.
Step 2A Prong 2: This judicial exception is not integrated into a practical application. No further additional elements are recited, see analysis of claim 2.
Step 2B: The claim does not contain significantly more than the judicial exception. No further additional elements are recited, see analysis of claim 2.
Claim 4
Step 1: A process, as above.
Step 2A Prong 1: The claim recites the same judicial exception as claim 1.
Step 2A Prong 2: This judicial exception is not integrated into a practical application. The claim further recites “storing the plurality of positive and negative query-context pairs for the unlabelled text and the corresponding context as a plurality of vector embeddings that includes multimodal vector” and “indexing the plurality of vector embeddings corresponding to the plurality of unlabelled texts,” however these limitations amount to insignificant post solution activity (MPEP 2106.05(g)).
Step 2B: The claim does not contain significantly more than the judicial exception. The storing and indexing limitations, in addition to being insignificant extra-solution activity, are also directed to the well-understood, routine, and conventional activity of storing and retrieving information in memory (MPEP 2106.05(d)(II)(iv) Versata Dev. Group, Inc. v. SAP Am., Inc., 793 F.3d 1306, 1334, 115 USPQ2d 1681, 1701 (Fed. Cir. 2015); OIP Techs., 788 F.3d at 1363, 115 USPQ2d at 1092-93).
Claim 5
Step 1: The claim is directed to a computer system and therefore is directed to the statutory category of machines.
Step 2A Prong 1: The claim recites, inter alia:
“retrieve a response from one or more data sources based on the input text…”; This limitation encompasses mentally retrieving a response from one or more data sources based on the input text, such as by reviewing one or more data sources and mentally choosing a response from a data source based on the input text.
“determining a context of each unlabelled text in the plurality of unlabelled texts…”; This limitation encompasses mentally determining a context of each unlabelled text, such as by mentally determining the subject matter of the unlabelled text.
“performing an augmentation operation on each unlabelled text and the context corresponding to each unlabelled text”; This limitation encompasses mentally performing an augmentation operation on each unlabelled text and the context corresponding to each unlabelled text, such as by mentally generating additional texts and contexts to augment each unlabelled text and corresponding context.
“generating a plurality of positive and negative query-context pairs for each unlabelled text and the corresponding context based on the augmentation operation…”; This limitation encompasses mentally generating a plurality of positive and negative query-context pairs for the unlabelled text and the corresponding context based on the augmentation operation, such as by mentally pairing positive queries with positive contexts and negative queries with negative contexts.
Step 2A Prong 2: This judicial exception is not integrated into a practical application. The claim further recites “receive an input text indicating a query made by a user”, however this limitation amounts to the insignificant extra-solution activity of mere data gathering (MPEP 2106.05(g)). The claim additionally recites a computer system comprising a memory and one or more processors configured to perform the method, “providing a plurality of unlabelled texts as training input data to the self-supervised Large Language retrieval model”, that the response is retrieved “using a self-supervised Large Language retrieval model,” that the context is determined “using one or more Artificial Intelligence (AI) techniques,” and that the plurality of positive and negative query-context pairs are generated “using the self-supervised Large Language retrieval model,” however, these limitations amount to mere instructions to apply an exception on a generic computer programmed with a generic class of computer algorithms (MPEP 2106.05(f)). The claim further recites “wherein the self-supervised Large language retrieval model is pretrained” and “configuring the self-supervised Large Language retrieval model to retrieve the response from the one or more data sources in response to each unlabelled text,” however, these limitations merely generally link the use of the judicial exception to the technological environment of model pre-training (MPEP 2106.05(h)).
Step 2B: The claim does not contain significantly more than the judicial exception. The “receive an input text” limitation, in addition to being insignificant extra-solution activity, is also directed to the well-understood, routine, and conventional activity of receiving or transmitting data over a network (MPEP 2106.05(d)(II)(i) buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355, 112 USPQ2d 1093, 1096 (Fed. Cir. 2014) (computer receives and sends information over a network)). Otherwise, the analysis at this step mirrors that of step 2A prong 2. As an ordered whole, the claim is directed to a mentally performable process of determining a context of unlabelled texts, performing an augmentation operation, and generating positive and negative query-context pairs for retrieving a response from one or more data sources. Nothing in the claim provides significantly more than this. As such, the claim is not patent eligible.
Claim 6
Step 1: A machine, as above.
Step 2A Prong 1: The claim recites, inter alia:
“…retrieve the response from the one or more data sources based on the labelled query-context pairs”; This limitation encompasses mentally retrieving a response from one or more data sources based on the labelled query-context pairs, such as by reviewing one or more data sources and mentally choosing a response from a data source based on the labelled query-context pairs.
Step 2A Prong 2: This judicial exception is not integrated into a practical application. The claim further recites “receive a plurality of test query-context representations such that each test query-context representation indicates a test query and a corresponding context of the test query,” however this limitation amounts to the insignificant extra-solution activity of mere data gathering (MPEP 2106.05(g)). The claim additionally recites “train the self-supervised Large Language retrieval model,” “provide labelled query-context pairs to the self-supervised Large Language retrieval model,” and “optimize the self-supervised Large Language retrieval model to… [perform the judicial exception],” however, these limitations merely generally link the use of the judicial exception to the technological environment of model training (MPEP 2106.05(h)).
Step 2B: The claim does not contain significantly more than the judicial exception. The “receive a plurality of test query-context representations” limitation, in addition to being insignificant extra-solution activity, is also directed to the well-understood, routine, and conventional activity of receiving or transmitting data over a network (MPEP 2106.05(d)(II)(i) buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355, 112 USPQ2d 1093, 1096 (Fed. Cir. 2014) (computer receives and sends information over a network)). Otherwise, the analysis at this step mirrors that of step 2A prong 2.
Claims 7
Step 1: A machine, as above.
Step 2A Prong 1: The claim recites, inter alia:
“evaluate a quadruplet loss function by evaluating a loss between the response retrieved by the self-supervised Large Language retrieval model and an expected response, wherein the quadruplet loss function is based on the positive and negative query-context pairs for each of the unlabelled text and the corresponding context”; This limitation encompasses mentally evaluating a quadruplet loss function by evaluating a loss between the response retrieved by the self-supervised Large Language retrieval model and an expected response.
Step 2A Prong 2: This judicial exception is not integrated into a practical application. No further additional elements are recited, see analysis of claim 5.
Step 2B: The claim does not contain significantly more than the judicial exception. No further additional elements are recited, see analysis of claim 5.
Claim 8
Step 1: A machine, as above.
Step 2A Prong 1: The claim recites the same judicial exception as claim 5.
Step 2A Prong 2: This judicial exception is not integrated into a practical application. The claim further recites “store the plurality of positive and negative query-context pairs for the unlabelled text and the corresponding context as a plurality of vector embeddings” and “index the plurality of vector embeddings corresponding to the plurality of unlabelled texts,” however these limitations amount to insignificant post solution activity (MPEP 2106.05(g)).
Step 2B: The claim does not contain significantly more than the judicial exception. The storing and indexing limitations, in addition to being insignificant extra-solution activity, are also directed to the well-understood, routine, and conventional activity of storing and retrieving information in memory (MPEP 2106.05(d)(II)(iv) Versata Dev. Group, Inc. v. SAP Am., Inc., 793 F.3d 1306, 1334, 115 USPQ2d 1681, 1701 (Fed. Cir. 2015); OIP Techs., 788 F.3d at 1363, 115 USPQ2d at 1092-93).
Claim 9
Step 1: A machine, as above.
Step 2A Prong 1: The claim recites, inter alia:
“identify a positive sample from the plurality of positive and negative query-context pairs when a semantic meaning associated with an augmented unlabelled text and the corresponding context corresponds to the semantic meaning associated with the unlabelled text and the corresponding context”; This limitation encompasses mentally identifying a positive sample from the plurality of positive and negative query-context pairs when a semantic meaning associated with an augmented unlabelled text and the corresponding context corresponds to the semantic meaning associated with the unlabelled text and the corresponding context.
“identify a negative sample from the plurality of positive and negative query-context pairs when a semantic meaning associated with an augmented unlabelled text and the corresponding context is different from the unlabelled text and the corresponding context”; This limitation encompasses mentally identifying a negative sample from the plurality of positive and negative query-context pairs when a semantic meaning associated with an augmented unlabelled text and the corresponding context is different from the unlabelled text and the corresponding context.
Step 2A Prong 2: This judicial exception is not integrated into a practical application. No further additional elements are recited, see analysis of claim 5.
Step 2B: The claim does not contain significantly more than the judicial exception. No further additional elements are recited, see analysis of claim 5.
Claim 10
Step 1: The claim recites a non-transitory computer readable media and therefore is directed to the statutory category of articles of manufacture.
Step 2A Prong 1: The claim recites, inter alia:
“retrieve a response from one or more data sources based on the input text…”; This limitation encompasses mentally retrieving a response from one or more data sources based on the input text, such as by reviewing one or more data sources and mentally choosing a response from a data source based on the input text.
“determining a context of each unlabelled text in the plurality of unlabelled texts…”; This limitation encompasses mentally determining a context of each unlabelled text, such as by mentally determining the subject matter of the unlabelled text.
“performing an augmentation operation on each unlabelled text and the context corresponding to each unlabelled text”; This limitation encompasses mentally performing an augmentation operation on each unlabelled text and the context corresponding to each unlabelled text, such as by mentally generating additional texts and contexts to augment each unlabelled text and corresponding context.
“generating a plurality of positive and negative query-context pairs for each unlabelled text and the corresponding context based on the augmentation operation…”; This limitation encompasses mentally generating a plurality of positive and negative query-context pairs for the unlabelled text and the corresponding context based on the augmentation operation, such as by mentally pairing positive queries with positive contexts and negative queries with negative contexts.
Step 2A Prong 2: This judicial exception is not integrated into a practical application. The claim further recites “receiving an input text indicating a query made by a user” however this limitation amounts to the insignificant extra-solution activity of mere data gathering (MPEP 2106.05(g)). The claim additionally recites a non-transitory computer readable media comprising instructions, when executed by a processor, cause the processor to perform the method, “providing a plurality of unlabelled texts as training input data to the self-supervised Large Language retrieval model,” that the response is retrieved “using a self-supervised Large Language retrieval model,” that the context is determined “using one or more Artificial Intelligence (AI) techniques,” and that the plurality of positive and negative query-context pairs are generated “using the self-supervised Large Language retrieval model,” however, these limitations amount to mere instructions to apply an exception on a generic computer programmed with a generic class of computer algorithms (MPEP 2106.05(f)). The claim further recites “wherein the self-supervised Large Language retrieval model is pretrained” and “configuring the self-supervised Large Language retrieval model to retrieve the response from the one or more data sources in response to each unlabelled text,” however, these limitations merely generally link the use of the judicial exception to the technological environment of model pre-training (MPEP 2106.05(h)).
Step 2B: The claim does not contain significantly more than the judicial exception. The “receiving an input text” limitation, in addition to being insignificant extra-solution activity, is also directed to the well-understood, routine, and conventional activity of receiving or transmitting data over a network (MPEP 2106.05(d)(II)(i) buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355, 112 USPQ2d 1093, 1096 (Fed. Cir. 2014) (computer receives and sends information over a network)). Otherwise, the analysis at this step mirrors that of step 2A prong 2. As an ordered whole, the claim is directed to a mentally performable process of determining a context of unlabelled texts, performing an augmentation operation, and generating positive and negative query-context pairs for retrieving a response from one or more data sources. Nothing in the claim provides significantly more than this. As such, the claim is not patent eligible.
Claim 11
Step 1: An article of manufacture, as above.
Step 2A Prong 1: The claim recites, inter alia:
“…retrieve the response from the one or more data sources based on the labelled query-context pairs”; This limitation encompasses mentally retrieving a response from one or more data sources based on the labelled query-context pairs, such as by reviewing one or more data sources and mentally choosing a response from a data source based on the labelled query-context pairs.
Step 2A Prong 2: This judicial exception is not integrated into a practical application. The claim further recites “receiving a plurality of test query-context representations such that each test query-context representation indicates a test query and a corresponding context of the test query” however this limitation amounts to the insignificant extra-solution activity of mere data gathering (MPEP 2106.05(g)). The claim additionally recites “training the self-supervised Large Language retrieval model,” “providing labelled query-context pairs to the self-supervised Large Language retrieval model,” and “optimizing the self-supervised Large Language retrieval model to… [perform the judicial exception],” however, these limitations merely generally link the use of the judicial exception to the technological environment of model training (MPEP 2106.05(h)).
Step 2B: The claim does not contain significantly more than the judicial exception. The “receive a plurality of test query-context representations” limitation, in addition to being insignificant extra-solution activity, are also directed to the well-understood, routine, and conventional activity of receiving or transmitting data over a network (MPEP 2106.05(d)(II)(i) buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355, 112 USPQ2d 1093, 1096 (Fed. Cir. 2014) (computer receives and sends information over a network)). Otherwise, the analysis at this step mirrors that of step 2A prong 2.
Claim 12
Step 1: An article of manufacture, as above.
Step 2A Prong 1: The claim recites, inter alia:
“evaluating a quadruplet loss function by evaluating a loss between the response retrieved by the self-supervised Large Language retrieval model and an expected response, wherein the quadruplet loss function is based on the positive and negative query-context pairs for each of the unlabelled text and the corresponding context”; This limitation encompasses mentally evaluating a quadruplet loss function by evaluating a loss between the response retrieved by the self-supervised Large Language retrieval model and an expected response.
Step 2A Prong 2: This judicial exception is not integrated into a practical application. No further additional elements are recited, see analysis of claim 10.
Step 2B: The claim does not contain significantly more than the judicial exception. No further additional elements are recited, see analysis of claim 10.
Claim 13
Step 1: An article of manufacture, as above.
Step 2A Prong 1: The claim recites the same judicial exception as claim 10.
Step 2A Prong 2: This judicial exception is not integrated into a practical application. The claim further recites “storing the plurality of positive and negative query-context pairs for the unlabelled text and the corresponding context as a plurality of vector embeddings that includes multimodal vector” and “indexing the plurality of vector embeddings corresponding to the plurality of unlabelled texts,” however these limitations amount to insignificant post solution activity (MPEP 2106.05(g)).
Step 2B: The claim does not contain significantly more than the judicial exception. The storing and indexing limitations, in addition to being insignificant extra-solution activity, are also directed to the well-understood, routine, and conventional activity of storing and retrieving information in memory (MPEP 2106.05(d)(II)(iv) Versata Dev. Group, Inc. v. SAP Am., Inc., 793 F.3d 1306, 1334, 115 USPQ2d 1681, 1701 (Fed. Cir. 2015); OIP Techs., 788 F.3d at 1363, 115 USPQ2d at 1092-93).
Claim 14
Step 1: An article of manufacture, as above.
Step 2A Prong 1: The claim recites, inter alia:
“identifying a positive sample from the plurality of positive and negative query-context pairs when a semantic meaning associated with an augmented unlabelled text and the corresponding context corresponds to the semantic meaning associated with the unlabelled text and the corresponding context”; This limitation encompasses mentally identifying a positive sample from the plurality of positive and negative query-context pairs when a semantic meaning associated with an augmented unlabelled text and the corresponding context corresponds to the semantic meaning associated with the unlabelled text and the corresponding context
“identifying a negative sample from the plurality of positive and negative query-context pairs when a semantic meaning associated with an augmented unlabelled text and the corresponding context is different from the unlabelled text and the corresponding context”; This limitation encompasses mentally identifying a negative sample from the plurality of positive and negative query-context pairs when a semantic meaning associated with an augmented unlabelled text and the corresponding context is different from the unlabelled text and the corresponding context
Step 2A Prong 2: This judicial exception is not integrated into a practical application. No further additional elements are recited, see analysis of claim 10.
Step 2B: The claim does not contain significantly more than the judicial exception. No further additional elements are recited, see analysis of claim 10.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 1, 5, and 10 are rejected under 35 U.S.C. 103 as being unpatentable over Guu et al. (NPL: “REALM: Retrieval-Augmented Language Model Pre-Training”) (hereinafter “Guu”) in view of Luo et al. (US20240412060) (hereinafter “Luo”) and further in view of Izacard et al. (NPL: “Unsupervised Dense Information Retrieval with Contrastive Learning”) (hereinafter “Izacard”).
Regarding claim 1, Guu discloses
“A method for retrieving a response using contrastive learning of contextual retrieval augmented generation, comprising:
receiving an input text indicating a query made by a user (Guu, Section 2. Open-domain question answering: “To measure a model’s ability to incorporate world knowledge, we need a downstream task where world knowledge is critical. Perhaps one of the most knowledge-intensive tasks in natural language processing is open-domain question answering (Open-QA): given a question x such as ‘What is the currency of the UK?’, a model must output the correct answer string y, ‘pound’); and
retrieving a response from one or more data sources based on the input text using a self-supervised Large Language retrieval model (Guu, Section 2. Open-domain question answering: “We focus on Open-QA systems that utilize a textual knowledge corpus Z as the knowledge source. Many of these systems employ a retrieval-based approach: given a question x, retrieve potentially relevant documents z from the corpus Z, and then extract an answer y from the documents... Our approach, REALM, is inspired by this paradigm and extends it to language model pre-training”; Examiner notes that the language model trained by REALM is a self-supervised retrieval model, as tertiary reference Izacard characterizes Guu as self-supervised learning for NLP (see Izacard Section 2 Related work, Self-supervised learning for NLP)),
wherein the self-supervised Large Language retrieval model is pre-trained (Guu, Section 2. Background, Language model pre-training) by:
providing a plurality of unlabelled texts as training input data to the self-supervised Large Language retrieval model (Guu, Figure 1: “Unlabeled text, from pre-training corpus X” input to “Neural Knowledge Retriever” and Section 2. Background, Language model pre-training: “Given an unlabeled pre-training corpus X (e.g., Wikipedia text), a training example (x,y) can be generated by randomly masking tokens in a sampled piece of text (e.g., x = “The [MASK] is the currency [MASK] the UK”; y = (“pound”, “of”))”);
determining, using one or more Artificial Intelligence (AI) techniques, a context for each unlabelled text in the plurality of unlabelled texts (Guu, Figure 1: “Retrieved document” from “Neural Knowledge Retriever”; Examiner notes that the retrieved document corresponds to a context of an unlabelled text because it is related to the query and provides additional information for understanding the query)…
configuring the self-supervised Large Language retrieval model to retrieve the response from the one or more data sources in response to each unlabelled text (Guu, Figure 1: “Answer” and Guu, 4.3. Implementation Details, Pre-training: “We pre-train for 200k steps on 64 Google Cloud TPUs, with a batch size of 512 and a learning rate of 3e-5, using BERT’s default optimizer. The document embedding step for the MIPS index is parallelized over 16 TPUs. For each example, we retrieve and marginalize over 8 candidate documents, including the null document ∅. We experiment with two choices of the pre-training corpus X: (1) Wikipedia, which is identical to the knowledge corpus Z, and (2) CC-News”).
Guu does not appear to explicitly disclose the further limitations of the claim.
However, Luo discloses “performing an augmentation operation on each… text and… [a] context corresponding to each… text” (Luo, [0004]: “The set of training data comprises, for each of a plurality of queries, respective query-document data and respective relevance judgement data. The query-document data for a query comprises data associated with a plurality of query-documents pairs for the query” and [0078]: “FIG. 9 illustrates a data perturbation operation in some embodiments of the invention. The data perturbation operation is arranged to facilitate a contrastive learning operation. As shown in FIG. 9, a query-document pair pi is processed by a data perturbation system 900 to generate augmented query-document pair(s), which includes two augmented query-document pairs pi′ and pi″ in some embodiments. The data perturbation operation may be performed for one or more of the query-document pairs of a query, and be performed for one or more of the queries. In some embodiments, the data perturbation system 900 is arranged to add noise to the query-document pair pi to generate the augmented query-document pair(s)”; Examiner notes that the query of query-document pair pi corresponds to a text, the document of query-document pair pi is the context corresponding to the text, and the data perturbation operation performed on query-document pair pi corresponds to “an augmentation operation,” which may be performed for one or more of the queries in the training data set) and
“generating a plurality of… query-context pairs for each… text and the corresponding context based on the augmentation operation…” (Luo, [0078]: “As shown in FIG. 9, a query-document pair pi is processed by a data perturbation system 900 to generate augmented query-document pair(s), which includes two augmented query-document pairs pi′ and pi″ in some embodiments”; Examiner notes that generating augmented query-document pairs pi’ and pi’’ by performing the augmentation operation on the original text and corresponding context (query-document pair pi) corresponds to “generating a plurality of query-context pairs for each text and the corresponding context based on the augmentation operation”).
Luo and the instant application both relate to machine learning and are analogous. It would have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified Guu with the teachings of Luo to include “performing an augmentation operation on each unlabelled text and the context corresponding to each unlabelled text” and “generating a plurality of… query-context pairs for each unlabelled text and the corresponding context based on the augmentation operation” and one would have been motivated to do so for the purpose of improving the model’s ability to generalize to unseen examples and learn a more uniform and robust representation, leading to more effective learning and better performance (see Luo, [0121]).
Neither Guu nor Luo appear to explicitly disclose that the plurality of query-context pairs are “positive and negative” or that they are generated using “the self-supervised Large Language retrieval model.”
However, Izacard discloses “generating… positive and negative query-context pairs… using… [a] retrieval model (Izacard 3.1.1 Contrastive learning: “A contrastive loss is used to learn by discriminating between documents. This loss compares either positive (from the same document) or negative (from different documents) pairs of document representations. Formally, given a query q with an associated positive document k+, and a pool of negative documents (ki)i=1..K, the contrastive InfoNCE loss is defined as: [see eq (1))] where τ is a temperature parameter. This loss encourages positive pairs to have high scores and negative pairs to have low scores. Another interpretation of this loss function is the following: given the query representation q, the goal is to recover, or retrieve, the representation k+ corresponding to the positive document, among all the negatives ki” and 4 Experiments: “In this section, we empirically evaluate our best retriever trained with contrastive learning, called Contriever (contrastive retriever), which uses MoCo with random cropping. We use a contrastive learning procedure that differs from ICT (Lee et al., 2019) mainly in three aspects. First, positive pairs are sampled using random cropping and tokens from each element of the pair are deleted with a probability of 10%. Second we use MoCo where negatives consists of elements from previous batches stored in a queue”; Examiner notes that (q,k+) in eq (1) corresponds to a positive query-context pair and (q,ki) corresponds to negative query-context pairs, which are generated by Contriever using MoCo with random cropping, and the Contriever model corresponds to a retrieval model).
Izacard and the instant application both relate to contrastive learning of information retrieval and are analogous. It would have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified the combination of Guu and Luo with the teachings of Izacard to include “generating a plurality of positive and negative query-context pairs for the unlabelled text and the corresponding context based on the augmentation operation using the self-supervised Large Language retrieval model,” and one would have been motivated to do so for the purpose of using contrastive learning for the pre-training of the language model to improve retrieval performance (see Izcard 4.3 Results, paragraph 2: “We observe that when used as pre training, contrastive learning leads to strong performance: contriever obtains the best results among dense bi encoder methods for the nDCG@10, and is state-of-the-art for the recall@100 (improving the average recall@100 from 65.0 to 67.1)”).
Regarding claim 5, Guu discloses
“A computer system, comprising: a memory; and one or more processors (Guu, Section 4.3. Implementation Details, Pre-training: “We pre-train for 200k steps on 64 Google Cloud TPUs, with a batch size of 512 and a learning rate of 3e-5, using BERT’s default optimizer”), configured to:
receive an input text indicating a query made by a user (Guu, Section 2. Open-domain question answering: “To measure a model’s ability to incorporate world knowledge, we need a downstream task where world knowledge is critical. Perhaps one of the most knowledge-intensive tasks in natural language processing is open-domain question answering (Open-QA): given a question x such as ‘What is the currency of the UK?’, a model must output the correct answer string y, ‘pound’); and
retrieve a response from one or more data sources based on the input text using a self-supervised Large Language retrieval model (Guu, Section 2. Open-domain question answering: “We focus on Open-QA systems that utilize a textual knowledge corpus Z as the knowledge source. Many of these systems employ a retrieval-based approach: given a question x, retrieve potentially relevant documents z from the corpus Z, and then extract an answer y from the documents... Our approach, REALM, is inspired by this paradigm and extends it to language model pre-training”; Examiner notes that the language model trained by REALM is a self-supervised retrieval model, as tertiary reference Izacard characterizes Guu as self-supervised learning for NLP (see Izacard Section 2 Related work, Self-supervised learning for NLP)),
wherein the self-supervised Large Language retrieval model is pre-trained (Guu, Section 2. Background, Language model pre-training) by:
providing a plurality of unlabelled texts as training input data to the self-supervised Large Language retrieval model (Guu, Figure 1: “Unlabeled text, from pre-training corpus X” input to “Neural Knowledge Retriever” and Section 2. Background, Language model pre-training: “Given an unlabeled pre-training corpus X (e.g., Wikipedia text), a training example (x,y) can be generated by randomly masking tokens in a sampled piece of text (e.g., x = “The [MASK] is the currency [MASK] the UK”; y = (“pound”, “of”))”);
determining a context of each unlabelled text in the plurality of unlabelled texts using one or more Artificial Intelligence (AI) techniques (Guu, Figure 1: “Retrieved document” from “Neural Knowledge Retriever”; Examiner notes that the retrieved document corresponds to a context of an unlabelled text because it is related to the query and provides additional information for understanding the query)…
configuring the self-supervised Large Language retrieval model to retrieve the response from the one or more data sources in response to each unlabelled text (Guu, Figure 1: “Answer” and Guu, 4.3. Implementation Details, Pre-training: “We pre-train for 200k steps on 64 Google Cloud TPUs, with a batch size of 512 and a learning rate of 3e-5, using BERT’s default optimizer. The document embedding step for the MIPS index is parallelized over 16 TPUs. For each example, we retrieve and marginalize over 8 candidate documents, including the null document ∅. We experiment with two choices of the pre-training corpus X: (1) Wikipedia, which is identical to the knowledge corpus Z, and (2) CC-News”).
Guu does not appear to explicitly disclose the further limitations of the claim.
However, Luo discloses “performing an augmentation operation on each… text and… context corresponding to each… text” (Luo, [0004]: “The set of training data comprises, for each of a plurality of queries, respective query-document data and respective relevance judgement data. The query-document data for a query comprises data associated with a plurality of query-documents pairs for the query” and [0078]: “FIG. 9 illustrates a data perturbation operation in some embodiments of the invention. The data perturbation operation is arranged to facilitate a contrastive learning operation. As shown in FIG. 9, a query-document pair pi is processed by a data perturbation system 900 to generate augmented query-document pair(s), which includes two augmented query-document pairs pi′ and pi″ in some embodiments. The data perturbation operation may be performed for one or more of the query-document pairs of a query, and be performed for one or more of the queries. In some embodiments, the data perturbation system 900 is arranged to add noise to the query-document pair pi to generate the augmented query-document pair(s)”; Examiner notes that the query of query-document pair pi corresponds to a text, the document of query-document pair pi is the context corresponding to the text, and the data perturbation operation performed on query-document pair pi corresponds to “an augmentation operation,” which may be performed for one or more of the queries in the training data set) and
“generating a plurality of… query-context pairs for each… text and the corresponding context based on the augmentation operation…” (Luo, [0078]: “As shown in FIG. 9, a query-document pair pi is processed by a data perturbation system 900 to generate augmented query-document pair(s), which includes two augmented query-document pairs pi′ and pi″ in some embodiments”; Examiner notes that generating augmented query-document pairs pi’ and pi’’ by performing the augmentation operation on the original text and corresponding context (query-document pair pi) corresponds to “generating a plurality of query-context pairs for each text and the corresponding context based on the augmentation operation”).
Luo and the instant application both relate to machine learning and are analogous. It would have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified Guu with the teachings of Luo to include “performing an augmentation operation on each unlabelled text and the context corresponding to each unlabelled text” and “generating a plurality of… query-context pairs for each unlabelled text and the corresponding context based on the augmentation operation” and one would have been motivated to do so for the purpose of improving the model’s ability to generalize to unseen examples and learn a more uniform and robust representation, leading to more effective learning and better performance (see Luo, [0121]).
Neither Guu nor Luo appear to explicitly disclose that the plurality of query-context pairs are “positive and negative” or that they are generated “using the self-supervised Large Language retrieval model.”
However, Izacard discloses “generating… positive and negative query-context pairs… using… [a] retrieval model (Izacard 3.1: “In this section, we describe our unsupervised training method…A contrastive loss is used to learn by discriminating between documents. This loss compares either positive (from the same document) or negative (from different documents) pairs of document representations. Formally, given a query q with an associated positive document k+, and a pool of negative documents (ki)i=1..K, the contrastive InfoNCE loss is defined as: [see eq (1))] where τ is a temperature parameter. This loss encourages positive pairs to have high scores and negative pairs to have low scores. Another interpretation of this loss function is the following: given the query representation q, the goal is to recover, or retrieve, the representation k+ corresponding to the positive document, among all the negatives ki” and 4 Experiments: “In this section, we empirically evaluate our best retriever trained with contrastive learning, called Contriever (contrastive retriever), which uses MoCo with random cropping. We use a contrastive learning procedure that differs from ICT (Lee et al., 2019) mainly in three aspects. First, positive pairs are sampled using random cropping and tokens from each element of the pair are deleted with a probability of 10%. Second we use MoCo where negatives consists of elements from previous batches stored in a queue”; Examiner notes that (q,k+) in eq (1) corresponds to a positive query-context pair and (q,ki) corresponds to negative query-context pairs, which are generated by Contriever using MoCo with random cropping, and the Contriever model corresponds to a retrieval model).
Izacard and the instant application both relate to contrastive learning of information retrieval and are analogous. It would have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified the combination of Guu and Luo with the teachings of Izacard to include “generating a plurality of positive and negative query-context pairs for the unlabelled text and the corresponding context based on the augmentation operation using the self-supervised Large Language retrieval model,” and one would have been motivated to do so for the purpose of using contrastive learning for the pre-training of the language model to improve retrieval performance (see Izcard 4.3 Results, paragraph 2: “We observe that when used as pre training, contrastive learning leads to strong performance: contriever obtains the best results among dense bi encoder methods for the nDCG@10, and is state-of-the-art for the recall@100 (improving the average recall@100 from 65.0 to 67.1)”).
Regarding claim 10, Guu discloses
“A non-transitory computer readable medium comprising instructions that when executed by a processor, cause the processor to: (Guu, Section 4.3. Implementation Details, Pre-training: “We pre-train for 200k steps on 64 Google Cloud TPUs, with a batch size of 512 and a learning rate of 3e-5, using BERT’s default optimizer”)
receive an input text indicating a query made by a user (Guu, Section 2. Open-domain question answering: “To measure a model’s ability to incorporate world knowledge, we need a downstream task where world knowledge is critical. Perhaps one of the most knowledge-intensive tasks in natural language processing is open-domain question answering (Open-QA): given a question x such as ‘What is the currency of the UK?’, a model must output the correct answer string y, ‘pound’); and
retrieve a response from one or more data sources based on the input text using a self-supervised Large Language retrieval model (Guu, Section 2. Open-domain question answering: “We focus on Open-QA systems that utilize a textual knowledge corpus Z as the knowledge source. Many of these systems employ a retrieval-based approach: given a question x, retrieve potentially relevant documents z from the corpus Z, and then extract an answer y from the documents... Our approach, REALM, is inspired by this paradigm and extends it to language model pre-training”; Examiner notes that the language model trained by REALM is a self-supervised retrieval model, as tertiary reference Izacard characterizes Guu as self-supervised learning for NLP (see Izacard Section 2 Related work, Self-supervised learning for NLP)),
wherein the self-supervised Large Language retrieval model is pre-trained (Guu, Section 2. Background, Language model pre-training) by:
providing a plurality of unlabelled texts as training input data to the self-supervised Large Language retrieval model (Guu, Figure 1: “Unlabeled text, from pre-training corpus X” input to “Neural Knowledge Retriever” and Section 2. Background, Language model pre-training: “Given an unlabeled pre-training corpus X (e.g., Wikipedia text), a training example (x,y) can be generated by randomly masking tokens in a sampled piece of text (e.g., x = “The [MASK] is the currency [MASK] the UK”; y = (“pound”, “of”))”);
determining a context of each unlabelled text in the plurality of unlabelled texts using one or more Artificial Intelligence (AI) techniques (Guu, Figure 1: “Retrieved document” from “Neural Knowledge Retriever”; Examiner notes that the retrieved document corresponds to a context of an unlabelled text because it is related to the query and provides additional information for understanding the query)…
configuring the self-supervised Large Language retrieval model to retrieve the response from the one or more data sources in response to each unlabelled text (Guu, Figure 1: “Answer” and Guu, 4.3. Implementation Details, Pre-training: “We pre-train for 200k steps on 64 Google Cloud TPUs, with a batch size of 512 and a learning rate of 3e-5, using BERT’s default optimizer. The document embedding step for the MIPS index is parallelized over 16 TPUs. For each example, we retrieve and marginalize over 8 candidate documents, including the null document ∅. We experiment with two choices of the pre-training corpus X: (1) Wikipedia, which is identical to the knowledge corpus Z, and (2) CC-News”).
Guu does not appear to explicitly disclose the further limitations of the claim.
However, Luo discloses “performing an augmentation operation on each… text and… context corresponding to each… text” (Luo, [0004]: “The set of training data comprises, for each of a plurality of queries, respective query-document data and respective relevance judgement data. The query-document data for a query comprises data associated with a plurality of query-documents pairs for the query” and [0078]: “FIG. 9 illustrates a data perturbation operation in some embodiments of the invention. The data perturbation operation is arranged to facilitate a contrastive learning operation. As shown in FIG. 9, a query-document pair pi is processed by a data perturbation system 900 to generate augmented query-document pair(s), which includes two augmented query-document pairs pi′ and pi″ in some embodiments. The data perturbation operation may be performed for one or more of the query-document pairs of a query, and be performed for one or more of the queries. In some embodiments, the data perturbation system 900 is arranged to add noise to the query-document pair pi to generate the augmented query-document pair(s)”; Examiner notes that the query of query-document pair pi corresponds to a text, the document of query-document pair pi is the context corresponding to the text, and the data perturbation operation performed on query-document pair pi corresponds to “an augmentation operation,” which may be performed for one or more of the queries in the training data set) and
“generating a plurality of… query-context pairs for each… text and the corresponding context based on the augmentation operation…” (Luo, [0078]: “As shown in FIG. 9, a query-document pair pi is processed by a data perturbation system 900 to generate augmented query-document pair(s), which includes two augmented query-document pairs pi′ and pi″ in some embodiments”; Examiner notes that generating augmented query-document pairs pi’ and pi’’ by performing the augmentation operation on the original text and corresponding context (query-document pair pi) corresponds to “generating a plurality of query-context pairs for each text and the corresponding context based on the augmentation operation”).
Luo and the instant application both relate to machine learning and are analogous. It would have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified Guu with the teachings of Luo to include “performing an augmentation operation on each unlabelled text and the context corresponding to each unlabelled text” and “generating a plurality of… query-context pairs for each unlabelled text and the corresponding context based on the augmentation operation” and one would have been motivated to do so for the purpose of improving the model’s ability to generalize to unseen examples and learn a more uniform and robust representation, leading to more effective learning and better performance (see Luo, [0121]).
Neither Guu nor Luo appear to explicitly disclose that the plurality of query-context pairs are “positive and negative” or that they are generated using “the self-supervised Large Language retrieval model.”
However, Izacard discloses “generating… positive and negative query-context pairs… using… [a] retrieval model (Izacard 3.1: “In this section, we describe our unsupervised training method…A contrastive loss is used to learn by discriminating between documents. This loss compares either positive (from the same document) or negative (from different documents) pairs of document representations. Formally, given a query q with an associated positive document k+, and a pool of negative documents (ki)i=1..K, the contrastive InfoNCE loss is defined as: [see eq (1))] where τ is a temperature parameter. This loss encourages positive pairs to have high scores and negative pairs to have low scores. Another interpretation of this loss function is the following: given the query representation q, the goal is to recover, or retrieve, the representation k+ corresponding to the positive document, among all the negatives ki” and 4 Experiments: “In this section, we empirically evaluate our best retriever trained with contrastive learning, called Contriever (contrastive retriever), which uses MoCo with random cropping. We use a contrastive learning procedure that differs from ICT (Lee et al., 2019) mainly in three aspects. First, positive pairs are sampled using random cropping and tokens from each element of the pair are deleted with a probability of 10%. Second we use MoCo where negatives consists of elements from previous batches stored in a queue”; Examiner notes that (q,k+) in eq (1) corresponds to a positive query-context pair and (q,ki) corresponds to negative query-context pairs, which are generated by Contriever using MoCo with random cropping, and the Contriever model corresponds to a retrieval model).
Izacard and the instant application both relate to contrastive learning of information retrieval and are analogous. It would have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified the combination of Guu and Luo with the teachings of Izacard to include “generating a plurality of positive and negative query-context pairs for the unlabelled text and the corresponding context based on the augmentation operation using the self-supervised Large Language retrieval model,” and one would have been motivated to do so for the purpose of using contrastive learning for the pre-training of the language model to improve retrieval performance (see Izcard 4.3 Results, paragraph 2: “We observe that when used as pre training, contrastive learning leads to strong performance: contriever obtains the best results among dense bi encoder methods for the nDCG@10, and is state-of-the-art for the recall@100 (improving the average recall@100 from 65.0 to 67.1)”).
Claims 2, 6, and 11 are rejected under 35 U.S.C. 103 as being unpatentable over Guu in view of Luo and Izacard, and further in view of Wu et al. (NPL: “Query-as-context Pre-training for Dense Passage Retrieval”) (hereinafter “Wu”).
Regarding claim 2, the rejection of claim 1 is incorporated. Guu as modified by Luo and Izacard does not appear to explicitly disclose the further limitations of the claim.
However, Wu discloses “training… [a] self-supervised Large Language retrieval model (Wu, 4.2 Fine-tuning, Datasets and Evaluation: “We fine-tune the pre trained coCondenser and CoT-MAE on MS MARCO passage ranking (Nguyen et al., 2016), TRECDeep Learning (DL) Track 2019 (Craswell et al., 2020a) and 2020 (Craswell et al., 2020b) tasks for evaluation”) by:
receiving a plurality of test query-context representations such that each test query-context representation indicates a test query and a corresponding context of the test query (Wu, 4.2 Fine-tuning, Datasets and Evaluation: “MS-MARCO(Nguyen et al., 2016) is a bench mark dataset that contains real user queries collected from Bing search and web pages, and includes approximately 8.8 million passages in total. The training set consists of around 500,000 annotated query-document pairs”);
providing labelled query-context pairs to the self-supervised Large Language retrieval model (Wu, 3.2 Fine-tuning: “The retrievers are trained with contrastive learning on the manually annotated passage-query pairs”); and
optimizing the self-supervised Large Language retrieval model to retrieve… [a] response from… one or more data sources based on the labelled query-context pairs (Wu, 3.2 Fine-tuning: “When training retriever 1, for query q+, the negative samples {p−} include in-batch negative passages and BM25mined hard negative passages. When training retriever 2, hard negatives are also mined using a well-trained retriever 1 and combined with the other negative passages to create the negative samples {p−}. Both stages are optimized using the InfoNCE loss”).
Wu and the instant application both relate to contrastive learning for response retrieval and are analogous. It would have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified the combination of Guu/Luo/Izacard with Wu to include “training the self-supervised Large Language retrieval model by: receiving a plurality of test query-context representations such that each test query-context representation indicates a test query and a corresponding context of the test query; providing labelled query-context pairs to the self-supervised Large Language retrieval model; and optimizing the self-supervised Large Language retrieval model to retrieve the response from the one or more data sources based on the labelled query-context pairs,” and one would have been motivated to do so for the purpose of fine-tuning the model on the downstream retrieval tasks to verify the effectiveness of the pre-training (see Wu, 3.2 Fine-tuning).
Regarding claim 6, the rejection of claim 5 is incorporated. Claim 6 is a system claim corresponding to method claim 2, and the rejection follows the same rationale given in the rejection of claim 2 above.
Regarding claim 11, the rejection of claim 10 is incorporated. Claim 11 is a computer readable medium claim corresponding to method claim 2, and the rejection follows the same rationale given in the rejection of claim 2 above.
Claim 3 is rejected under 35 U.S.C. 103 as being unpatentable over Guu in view of Luo, Izacard, and Wu, and further in view of Xu et al. (NPL: “HSimCSE: Improving Contrastive Learning of Unsupervised Sentence Representation with Adversarial Hard Positives and Dual Hard Negatives”) (hereinafter “Xu”).
Regarding claim 3, the rejection of claim 2 is incorporated. Guu as modified by Luo, Izacard, and Wu further discloses “evaluating a… loss function by evaluating a loss between the response retrieved by the self-supervised Large Language retrieval model and an expected response, wherein the… loss function is based on the positive and negative query-context pairs for each of the unlabelled text and the corresponding context (Izacard, 3.1.1 Contrastive learning: “A contrastive loss is used to learn by discriminating between documents. This loss compares either positive (from the same document) or negative (from different documents) pairs of document representations. Formally, given a query q with an associated positive document k+, and a pool of negative documents (ki)i=1..K, the contrastive InfoNCE loss is defined as: [see eq (1)] where τ is a temperature parameter. This loss encourages positive pairs to have high scores and negative pairs to have low scores”).
Izacard and the instant application both relate to contrastive learning of information retrieval and are analogous. It would have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified the combination of Guu/Luo/Wu to with the teachings of Izacard include “evaluating a… loss function by evaluating a loss between the response retrieved by the self-supervised Large Language retrieval model and an expected response, wherein the… loss function is based on the positive and negative query-context pairs for each of the unlabelled text and the corresponding context” and one would have been motivated to do so for the purpose of encouraging positive pairs to have high scores and negative pairs to have low scores (see Izacard, 3.1.1 Contrastive learning).
Guu as modified by Luo, Izacard, and Wu does not appear to explicitly disclose that the loss function is a “quadruplet loss function.”
However, Xu discloses “evaluating a quadruplet loss function…wherein the quadruplet loss function is based on… positive and negative… pairs” (Xu, III. Method, D. Objective Function: “Finally, based on the existing contrastive loss (i.e., NT-Xent in SimCSE), we additionally use a quadruplet loss function to train the sentence encoder. The basic idea of the quadruplet loss is to minimize the distance between the anchor and the adversarial hard positive samples and maximize the distance between the anchor sample and the closest local hard negative sample and the mixed global hard negative sample. Specifically, the quadruplet loss function takes one anchor sample, one hard positive sample, and two hard negative samples as input, which is defined as follows: [see eq (12)] where n is the batch size, C is a margin between positive and negative pairs and η is the hyper-parameter used to control the importance of different negative samples”).
Xu and the instant application both relate to contrastive learning and are analogous. It would have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified the combination of Guo/Luo/Izacard/Wu with the teachings of Xu such that the loss function is a quadruplet loss function, and one would have been motivated to do so for the purpose of obtaining high-quality representations of sentences in an unsupervised manner to apply to the downstream application of question answering (see Xu, I. Introduction and II. Overview, A. Problem Formulation).
Claims 7 and 12 are rejected under 35 U.S.C. 103 as being unpatentable over Guu in view of Luo and Izacard, and further in view of Xu.
Regarding claim 7, the rejection of claim 5 is incorporated. Guu as modified by Luo and Izacard further discloses “evaluate a… loss function by evaluating a loss between the response retrieved by the self-supervised Large Language retrieval model and an expected response, wherein the… loss function is based on the positive and negative query-context pairs for each of the unlabelled text and the corresponding context (Izacard, 3.1.1 Contrastive learning: “A contrastive loss is used to learn by discriminating between documents. This loss compares either positive (from the same document) or negative (from different documents) pairs of document representations. Formally, given a query q with an associated positive document k+, and a pool of negative documents (ki)i=1..K, the contrastive InfoNCE loss is defined as: [see eq (1)] where τ is a temperature parameter. This loss encourages positive pairs to have high scores and negative pairs to have low scores”).
Izacard and the instant application both relate to contrastive learning of information retrieval and are analogous. It would have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified the combination of Guu and Luo with the teachings of Izacard to include the one or more processors being further configured to “evaluate a… loss function by evaluating a loss between the response retrieved by the self-supervised Large Language retrieval model and an expected response, wherein the… loss function is based on the positive and negative query-context pairs for each of the unlabelled text and the corresponding context” and one would have been motivated to do so for the purpose of encouraging positive pairs to have high scores and negative pairs to have low scores (see Izacard, 3.1.1 Contrastive learning).
Guu as modified by Luo and Izacard does not appear to explicitly disclose that the loss function is a “quadruplet loss function.”
However, Xu discloses “evaluate a quadruplet loss function…wherein the quadruplet loss function is based on… positive and negative… pairs” (Xu, III. Method, D. Objective Function: “Finally, based on the existing contrastive loss (i.e., NT-Xent in SimCSE), we additionally use a quadruplet loss function to train the sentence encoder. The basic idea of the quadruplet loss is to minimize the distance between the anchor and the adversarial hard positive samples and maximize the distance between the anchor sample and the closest local hard negative sample and the mixed global hard negative sample. Specifically, the quadruplet loss function takes one anchor sample, one hard positive sample, and two hard negative samples as input, which is defined as follows: [see eq (12)] where n is the batch size, C is a margin between positive and negative pairs and η is the hyper-parameter used to control the importance of different negative samples”).
Xu and the instant application both relate to contrastive learning and are analogous. It would have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified the combination of Guo/Luo/Izacard with the teachings of Xu such that the loss function is a quadruplet loss function, and one would have been motivated to do so for the purpose of obtaining high-quality representations of sentences in an unsupervised manner to apply to the downstream application of question answering (see Xu, I. Introduction and II. Overview, A. Problem Formulation).
Regarding claim 12, the rejection of claim 10 is incorporated. Claim 12 is a computer readable medium claim corresponding to system claim 7, and the rejection of claim 12 follows the same rationale given for claim 7 above.
Claims 4, 8, and 13 are rejected under 35 U.S.C. 103 as being unpatentable over Guu in view of Luo and Izacard, and further in view of Chen et al. (NPL: “MuRAG: Multimodal Retrieval-Augmented Generator for Open Question Answering over Images and Text”) (hereinafter “Chen”).
Regarding claim 4, the rejection of claim 1 is incorporated. Guu as modified by Luo and Izacard discloses “the plurality of positive and negative query-context pairs for the unlabelled text and the corresponding context” (see rejection of claim 1), but does not appear to explicitly disclose the further limitations of the claim.
However, Chen discloses “storing… [a] plurality of positive and negative query-context pairs… as a plurality of vector embeddings that includes multimodal vector (Chen 3.4 Fine-tuning: “As depicted in the lower part of Figure 4, the input query for the downstream task is a text question xq, and the memory M containing (image, text) pairs (mI,mT).1 The Top-K retrievals {(
m
1
I
,
m
1
T
),··· ,(
m
k
I
,
m
k
T
)} are leveraged to generate the answer a… The in-batch memory MB is constructed in the following way: the k-th example in the dataset is represented as [see equation] where m represents the positive (image, text) source, and m̄ represents the hard negative (image, text) source provided by the dataset2. For a batch with B examples, we assemble all the associated positive and negative knowledge source as our in-batch memory MB = [see equation]… After in-batch training, we encode all available cross-modal pairs”; Examiner notes that the positive and negative image text pairs correspond to “a plurality of positive and negative query-context pairs” which are encoded as cross-modal vector embeddings in in-batch memory) and indexing the plurality of vector embeddings corresponding to… [a] plurality of… texts” (Chen 3.4 Fine-tuning, Fixed-Retrieval Training: “After in-batch training, we encode all available cross-modal pairs, and index these encodings for fast MIPS retrieval”; Examiner notes that each of the indexed embeddings corresponds to a text since each embedding is an image, text pair).
Chen and the instant application both relate to retrieval-augmented generation and are analogous. It would have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified the combination of Guu/Luo/Izacard with the teachings of Chen to include “storing the plurality of positive and negative query-context pairs for the unlabelled text and the corresponding context as a plurality of vector embeddings that includes multimodal vector; and indexing the plurality of vector embeddings corresponding to the plurality of unlabelled texts,” and one would have been motivated to do so for the purpose of achieving fast retrieval (see Chen 3.4 Fine-tuning, Fixed-Retrieval Training).
Regarding claim 8, the rejection of claim 5 is incorporated. Guu as modified by Luo and Izacard discloses “the plurality of positive and negative query-context pairs for the unlabelled text and the corresponding context” (see rejection of claim 5), but does not appear to explicitly disclose the further limitations of the claim.
However, Chen discloses “store… [a] plurality of positive and negative query-context pairs… as a plurality of vector embeddings (Chen 3.4 Fine-tuning: “As depicted in the lower part of Figure 4, the input query for the downstream task is a text question xq, and the memory M containing (image, text) pairs (mI,mT).1 The Top-K retrievals {(
m
1
I
,
m
1
T
),··· ,(
m
k
I
,
m
k
T
)} are leveraged to generate the answer a… The in-batch memory MB is constructed in the following way: the k-th example in the dataset is represented as [see equation] where m represents the positive (image, text) source, and m̄ represents the hard negative (image, text) source provided by the dataset2. For a batch with B examples, we assemble all the associated positive and negative knowledge source as our in-batch memory MB = [see equation]… After in-batch training, we encode all available cross-modal pairs”; Examiner notes that the positive and negative image text pairs correspond to “a plurality of positive and negative query-context pairs” which are encoded as cross-modal vector embeddings in in-batch memory) and index the plurality of vector embeddings corresponding to… [a] plurality of unlabelled texts” (Chen 3.4 Fine-tuning, Fixed-Retrieval Training: “After in-batch training, we encode all available cross-modal pairs, and index these encodings for fast MIPS retrieval”; Examiner notes that each of the indexed embeddings corresponds to a text since each embedding is an image, text pair).
Chen and the instant application both relate to retrieval-augmented generation and are analogous. It would have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified the combination of Guu/Luo/Izacard with the teachings of Chen to include the one or more processors being further configured to “store the plurality of positive and negative query-context pairs for the unlabelled text and the corresponding context as a plurality of vector embeddings; and index the plurality of vector embeddings corresponding to the plurality of unlabelled texts,” and one would have been motivated to do so for the purpose of achieving fast retrieval (see Chen 3.4 Fine-tuning, Fixed-Retrieval Training).
Regarding claim 13, the rejection of claim 10 is incorporated. Claim 13 is a computer readable medium claim corresponding to method claim 4, and the rejection of claim 13 follows the same rationale given for claim 4 above.
Claims 9 and 14 are rejected under 35 U.S.C. 103 as being unpatentable over Guu in view of Luo and Izacard, and further in view of Ge et al. (NPL: “Robust Contrastive Learning Using Negative Samples with Diminished Semantics”) (hereinafter “Ge”).
Regarding claim 9, the rejection of claim 5 is incorporated. Guu as modified by Luo and Izacard discloses “an augmented unlabelled text and the corresponding context,” and “the unlabelled text and the corresponding context” (see rejection of claim 5), but does not appear to explicitly disclose the further limitations of the claim.
However, Ge discloses
“identify a positive sample from… [a] plurality of positive and negative query-context pairs when a semantic meaning associated with an augmented… [input] corresponds to the semantic meaning associated with the… [original input] (Ge 2.1 Contrastive learning with non-semantic negatives: “We use zi and zp to denote the representations of the query sample xi and a positive sample xp generated from the same input image with augmentations that preserve semantics”; Examiner notes that zi and zp correspond to a query-context pair, and zp is a positive sample from the pair as it is an augmented version of the input that shares a semantic meaning with the original input); and
identify a negative sample from… [a] plurality of positive and negative query-context pairs when a semantic meaning associated with an augmented… [input] is different from the [original input] (Ge 2.1 Contrastive learning with non-semantic negatives: “zns is the representation of the proposed negative sample xns which contains particular non-semantic features of the input image with the semantic part weakened”; Examiner notes that zi and zns correspond to a query-context pair, and zns is a negative sample from the pair as it is an augmented version of the input that has a different semantic meaning than the original input (see also Figure 1)).
Ge and the instant application both relate to contrastive learning and are analogous. It would have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified the combination of Guu/Luo/Izacard with the teachings of Ge to include the processor being further configured to “identify a positive sample from the plurality of positive and negative query-context pairs when a semantic meaning associated with an augmented unlabelled text and the corresponding context corresponds to the semantic meaning associated with the unlabelled text and the corresponding context; and identify a negative sample from the plurality of positive and negative query-context pairs when a semantic meaning associated with an augmented unlabelled text and the corresponding context is different from the unlabelled text and the corresponding context,” and one would have been motivated to do so for the purpose of achieving improved generalization, especially under out-of-domain settings (see Ge, Abstract).
Regarding claim 14, the rejection of claim 10 is incorporated. Claim 14 is a computer readable medium claim corresponding to system claim 9, and the rejection of claim 14 follows the same rationale given in the rejection of claim 9 above.
Response to Arguments
Applicant’s arguments filed August 31, 2026, regarding the rejections under 35 U.S.C. 101 (Remarks, pages 10-19) have been fully considered but they are not persuasive.
Applicant argues regarding Step 2A, Prong One that the limitation of “determining a context of each unlabelled text using one or more AI techniques” is not a mental process because humans cannot practically “process millions of training instances; derive contextual representations for each training instance; [and] maintain contextual relationships across high-dimensional embedding spaces”. Examiner submits that none of these above-mentioned steps are required by the claim language. Under the broadest reasonable interpretation of the limitation, “determining a context of each unlabelled text” encompasses mentally determining a context of each unlabelled text, such as by observing the word “baseball” in an unlabelled text and mentally determining a context to be “sports”. The fact that the determining is performed “using one or more AI techniques” merely amounts to applying the exception with generic computer algorithms (MPEP 2106.05(f)). Applicant further contends that “performing an augmentation operation on each unlabelled text and the context corresponding to each unlabelled text; and generating positive and negative query-context pairs based on the augmentation operation” is not a mental process because a human cannot realistically “generate thousands or millions of contextual augmentations; preserve semantic relationships algorithmically; create contrastive-learning datasets; [and] maintain embedding relationships during retrieval-model training”. Examiner submits, as above, that none of these steps are required by the claim language. Thus, the limitation of “performing an augmentation operation on each unlabelled text and the context corresponding to each unlabelled text; and generating positive and negative query-context pairs based on the augmentation operation” encompasses mentally performing augmentation and mentally generating pairs. Applicant additionally argues that the retrieval step is not a mental process because it “occurs using vector embeddings and indexed retrieval structures.” However, the claim does not require that the retrieval occur using vector embeddings and indexed retrieval structures, it merely states “retrieving a response from one or more data sources based on the input text”, which encompasses mentally retrieving the response. Finally, Applicant argues that the rejection has “failed to provide any clear analysis besides generic states for Applicant to identify why claims 1, 5, and 10 recite alleged abstract ideas”. Examiner respectfully disagrees. Examiner has identified each limitation that recites a mental process and explained why each limitation is mentally performable, thus meeting the guidelines set forth in MPEP § 2106.04(a)(1).
Applicant argues regarding Step 2A, Prong Two that the claimed features of independent claim 1 integrate the abstract idea into a practical application. Applicant asserts that the specification describes that the computer system can be used in different use cases such as “code generation, conversational systems, information retrieval systems, sentiment analysis, text generation” and that these use cases “are clearly practical application[s] where the claimed invention can be implemented.” However, none of these use cases are recited in the claims, and even assuming, arguendo, that they were, merely generally linking the judicial exception to a particular field of use does not integrate a judicial exception into a practical application (MPEP 2106.04(d)(I)). Applicant further argues that the claimed architecture “provides technological improvements expressly described in the Specification: improved context retrieval; reduced hallucination; improved scalability; improved retrieval efficiency; improved generalization”. Examiner submits that these alleged improvements are set forth in a conclusory manner, and there is no clear nexus between one or more additional elements of the claim and any of these alleged improvements. Applicant suggests that this practical integration is analogous to the claims in Diamond v. Diehr and Example 36 of the Subject Matter Eligibility Examples, which were allegedly found eligible because they used computed information to control system operation. Examiner disagrees with this characterization. The claims in Diamond v. Diehr recite additional elements such as the use of a thermocouple to monitor temperature inside a rubber mold that integrate the judicial exception into a practical application. Claims 2 and 3 in Example 36 also recite additional elements that integrate the judicial exception into a practical application, such as “reconstruct the 3-D coordinates of the item of inventory from multiple overlapping images obtained from the camera array and prior knowledge of the location and field of view of the camera(s)”. No such additional elements are present here — as the additional elements of the instant claims amount to generic computer components, mere data gathering, and generically recited model training.
Applicant argues regarding Step 2B that independent claim 1 amounts to “significantly more” than an abstract idea because it represents a non-conventional and non-generic arrangement of elements and integrates into systems “for example information retrieval, sentiment analysis, text generation, code generation and conversational systems.” Examiner submits that no such system integration is recited in independent claim 1, nor does independent claim 1 contain any additional elements beyond the judicial exception other than what is well-understood, routine, and conventional. Thus, Examiner maintains the rejection under 35 U.S.C. 101.
Applicant’s arguments filed August 31, 2026, regarding the rejections under 35 U.S.C. 103 (Remarks, pages 19-24) have been fully considered but they are not persuasive.
Applicant argues that Guu does not teach the limitation of “determining, using one or more AI techniques, a context of each unlabelled text in the plurality of unlabelled texts”, contending that a retrieved document is not equivalent to a determined context. Examiner respectfully disagrees. Applicant’s specification does not provide a limiting definition for the word “context”; thus, Examiner is interpreting “context” by its plain English definition: “the parts of a discourse that surround a word, phrase, or passage and that help to explain its meaning” (see Merriam-Webster dictionary). The retrieved document shown in Figure 1 of Guu provides surrounding words that help to explain the meaning of the unlabeled text; therefore, the retrieved document amounts to a context of the unlabeled text, the Neural Knowledge Retriever that retrieves the document corresponds to “one or more AI techniques”, and the act of choosing which document to retrieve from the textual knowledge corpus amounts to a “determination”. Therefore, Guu teaches the limitation of “determining, using one or more AI techniques, a context of each unlabelled text in the plurality of unlabelled texts”.
Applicant further contends that Luo does not teach the limitations of “performing an augmentation operation on each unlabelled text and the context corresponding to each unlabelled text; and generating query-context pairs based on the augmentation operation” because Luo “merely teaches modifying an existing query-document pair”. Examiner respectfully disagrees. See for example Fig. 9 and paragraph [0078] of Luo: augmentation is performed on query-document pair pi (text and corresponding context) to generate augmented query-documents pairs
p
i
'
and
p
i
'
'
(query-context pairs). Thus, Luo doesn’t merely teach “modifying an existing query-document pair” — Luo teaches generating multiple query-context pairs as a result of the augmentation operation on the original text and corresponding context.
Applicant additionally argues that Izacard does not teach “generating positive and negative query-context pairs” because the “positive and negative relationships already exist in the dataset”. Examiner respectfully disagrees. Izacard discloses “First, positive pairs are sampled using random cropping and tokens for each element of the pair are deleted with a probability of 10%. Second we use MoCo where negatives consists of elements from previous batches stored in a queue”. Thus, the act of forming a positive pair (q, k+) by pairing a query with an augmented version of itself and forming negative pairs (q, ki) by pairing the query with elements from previous batches for use in the loss function of eq (1) to perform contrastive learning amounts to “generating positive and negative query-context pairs”.
In response to applicant's argument that the examiner's conclusion of obviousness is based upon improper hindsight reasoning, it must be recognized that any judgment on obviousness is in a sense necessarily a reconstruction based upon hindsight reasoning. But so long as it takes into account only knowledge which was within the level of ordinary skill at the time the claimed invention was made, and does not include knowledge gleaned only from the applicant's disclosure, such a reconstruction is proper. See In re McLaughlin, 443 F.2d 1392, 170 USPQ 209 (CCPA 1971).
Applicant argues that a further indication of non-obviousness is that the references attempt to solve different problems, and that a POSITA would not have been motivated to combine these references. Examiner respectfully disagrees. Each of the prior art references is in the same field of the inventor’s endeavor, and Examiner has provided a motivation to combine for each of the references, none of which Applicant has specifically refuted, thus there exists a prima facie case for obviousness.
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
Any inquiry concerning this communication or earlier communications from the examiner should be directed to GWYNEVERE A DETERDING whose telephone number is (571)272-7657. The examiner can normally be reached Mon-Fri. 9am-5pm.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Kamran Afshar can be reached at (571) 272-7796. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/G.A.D./Examiner, Art Unit 2125
/KAMRAN AFSHAR/Supervisory Patent Examiner, Art Unit 2125
